| # Methodology: CodeMixCheck v2.2.1 |
|
|
| ## Task design |
|
|
| CodeMixCheck treats fact verification as **claim-level veracity classification with an explicit evidence state**. Each synthetic claim (`proposition`) is assigned a veracity state in a simulated world, and is then rendered through different frames, post formats and levels of noise. This separates: |
|
|
| - *what* is claimed (the proposition, with its veracity), |
| - *how* it is framed (forwarded message / official attribution / hearsay / report), |
| - *how noisy* the surface text is. |
|
|
| All four labels occur under every frame family. This is intended to stop a classifier from equating hearsay framing with falsehood. |
|
|
| `evidence_explanation` is a short justification (mostly templated) that states the simulated evidence. **No retrievable evidence corpus is included.** RAG experiments therefore need an external or separately built evidence store, and that store must stay consistent with the simulated world. |
|
|
| ## AI-assisted generation |
|
|
| | component | what is recorded | what is not recorded | |
| |---|---|---| |
| | template rows | `provenance = template_generated`, batch, frame, format, noise | the generating script, model or chat assistant used to write templates/entity pools, prompts, dates | |
| | curated rows | `human_curated_llm_assisted`, `news_sourced_and_llm_assisted` | which LLM, which prompts, which sources, how much text the LLM wrote versus the author | |
| | silver sentiment | `model_annotated_v1.3` | the model behind "v1.3" | |
|
|
| The project used Claude and ChatGPT during dataset construction. Because no per-row or per-batch attribution was stored, this release labels the generator `not_recorded` rather than assigning a model to any record. |
|
|
| ## Splitting |
|
|
| - Fact verification: train 5,337 / validation 662 / synthetic test 661 / curated test 344. The synthetic splits are drawn from the same generator. The curated test is held out by origin. |
| - Splits are grouped by `content_group` and `template_group`. Neither spans two splits. |
| - v2.1 removed template-family leakage by moving 22 rows. |
| - The curated test carries `cv_fold` (0–4; folds of 69/69/69/69/68 rows) for cross-validated evaluation. |
|
|
| ## Recommended evaluation protocol |
|
|
| 1. Report macro-F1 and per-class F1 separately for `test_synthetic` and (when available) `test_curated`. |
| 2. For `test_curated`, also report results per `eval_subset`, because the label distributions differ sharply (hearsay = 100% Unverifiable; health/lifestyle = 90% True). |
| 3. Report a *strict* `test_synthetic` score that excludes rows flagged D01/D02 (cross-split same-claim or near-duplicate rows). |
| 4. Do not tune on `test_*`. Sentiment and abuse have very small validation sets, so expect high variance and report confidence intervals. |
| 5. Do not train on `silver_unverified` and then describe the result as trained on gold. |
|
|
